Invalid read of size 4 in mkl_serv_malloc

Invalid read of size 4 in mkl_serv_malloc

erling_andersen's picture

We have a crash in MKLon Intel 64x86 using Intel 13.1.3. It is both from valgrind and gdb. We use

     mkl_disable_fast_mm();

So it seems the buffer stuff should never be allocated.  Well, I am not sure but the buffer stuff seems in use.

Any suggestions, comments?

[My application which is a .so linked with a static version version of MKL into a MATLAB mex file. A MATLAB mex file is also a so it is called from MATLAB. Now MATLAB is is linked with MKL but another version and the shared lib. The crash above only pops up if called from MATLAB. Maybe this is a special situation you have not considered.]

The stacj trace should come here but your spam filer prevented me from adding it.

8 posts / 0 new
Last post
For more complete information about compiler optimizations, see our Optimization Notice.
erling_andersen's picture

I cannot paste the stack trace becuse of the spam filer but it is attached

Attachments: 

AttachmentSize
Download trace.txt2.32 KB
erling_andersen's picture

We are using   composer_xe_2013.5.192

Dmitry Baksheev (Intel)'s picture

Hi,

Could you print and paste here the parameters you pass to cblas_dgemm in MSK_INTEL_P4_mathenv_gemm (mathenvcpu.c:2293) ? That would help us to reproduce the issue.

Thanks
Dima

erling_andersen's picture

I have spend more time rsearching the issue. It seems

  • The problem is independent of the matrix sizes.
  • The problem occurs in the first call to dgemm. I moved the call to dgemm to begning of my program. To reduce the risk I overwrote the your data. It still occurs.

So the problem seems to be related to initialization of perhaps your intial data structures. Does that help. We link static as I mentioned.

I might be able to create an example but it will require you can run MATLAB r2013a on linux64x86. Is that of interest?

erling_andersen's picture

I also tried linking with  libmkl_sequential.so and then the problem also goes away.

However, I REALLY HATE that solution because distributing more shared libraries is a major PAIN. Also I had to set LD_PRELOAD to get it working which is major PAIN too.

   

erling_andersen's picture

I have reported a test case in  http://software.intel.com/en-us/forums/topic/486842

Chao Y (Intel)'s picture

Thanks.  I will check in that one.

Login to leave a comment.